Training-time sensitivity to the simulation–experiment weighting

Determine how training-time sensitivity to the simulation–experiment loss weight \(\lambda\) affects the performance of joint simulation–experiment learning beyond the single value \(\lambda=0.5\) used in the study.

Background

The study trains the Joint protocol using an equal weighting of simulation and experimental risks, λ=0.5\lambda=0.5, and separately evaluates fixed checkpoints under different evaluation weightings. However, changing the evaluation weighting does not reveal how the optimization process itself responds to different training-time scalarizations of the two domain objectives. The authors explicitly identify this training-time sensitivity as unexplored, leaving unresolved how alternative values of λ\lambda would influence balanced performance and the retention–adaptation trade-off.

References

Joint is trained only with $\lambda=0.5$. Appendix~\ref{app:lambda} varies the evaluation weighting with fixed checkpoints, but training-time sensitivity to $\lambda$ remains unexplored.

— Sim+Real: Joint Simulation - Experiment Training Improves Balanced Prediction in Physical Systems  (2610.01974 - Rautela et al., 1 Oct 2026) in Section Limitations